feat(ceus,segment): 任务书 02 第二阶段——分割和后半段算法

- ceus 迁入 polygons、dark_region、reference_region、tic_fitting、tic、anchor_frames、 perfusion_metrics、trajectory、tic_curves、tic_plot、overlay，旧测试照搬（改 import 与配置读取）。 - 对外两段接口：prepare_segmentation（第 1–2 步）与 quantify_case（第 4–9 步）。 - 第 3 条：删掉 D1 取医生标注时刻的模式，D1 只从拟合曲线算。 - 第 4 条：F1/F2 取视频第 20−X、60−X 秒（X 为视频第 0 秒时已推注的秒数），落在视频外记未定位。 - 绿框融合权重、紫圈颜色阈值照搬；关键帧容差窗口的出处写进 ceus.yaml 注释。 - segment：同一进程调用 MedSAM2，设置写进 ceus.yaml 的 segment 段，全部必填；gpu 测试真跑通过。 - 依赖加 scipy、matplotlib、medsam2（可编辑安装，不编译 CUDA 扩展）；uv 只解析 linux x86_64。 - 导入契约：只有 segment 引用 torch、sam2；segment 不引用其他 cadx 模块。 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-16 05:07:02 -04:00
parent 9a916c65a5
commit 150eb5f6ae
35 changed files with 10655 additions and 505 deletions
@@ -0,0 +1,168 @@
+"""ceus 对外两段接口的端到端测试：合成一段分屏造影视频，第二段喂合成掩膜（不调分割模型）。
+
+视频 400×200、10 fps、40 秒（400 帧），分屏线在 x=200（这一列全黑）。造影画面在右半：病灶是整帧
+[250, 369] × [60, 159] 的方块，医生的绿框是它的空心描边；病灶内部亮度按 Lognormal 团注曲线从
+基线 40 升到约 170 再回落，内部另有一块一直发黑的囊（暗区）；每帧加一点随机噪声，免得相邻帧
+被当成重复帧。左半是不变的灰阶画面。
+
+手算的预期：
+- 绿框在右半 → 造影画面是右半，分割那一半是左半 [0, 200)；
+- 绿框挪到左半局部坐标：x = 250 − 200 = 50 到 369 − 200 = 169；
+- 推注时间 X = 5 秒 → F1 在视频第 20 − 5 = 15 秒（第 150 帧），F2 在第 60 − 5 = 55 秒，
+  超过 40 秒的视频末尾，未定位。
+"""
+
+import json
+import math
+from pathlib import Path
+
+import cv2
+import numpy as np
+import pytest
+
+from cadx.ceus import (
+    ANCHOR_SLOTS,
+    CeusConfig,
+    LognormalParameters,
+    NoValidContourError,
+    SegmentationInput,
+    evaluate_lognormal,
+    keyframe_images,
+    pack_tic_curves,
+    pack_trajectory,
+    prepare_segmentation,
+    quant_summary,
+    quantify_case,
+    render_case_overlay,
+    render_tic_png,
+    tic_plot_markers,
+    unpack_tic_curves,
+    unpack_trajectory,
+)
+
+_WIDTH, _HEIGHT, _SPLIT = 400, 200, 200
+_FPS, _FRAMES = 10.0, 400
+_LESION = (250, 60, 369, 159)  # 整帧坐标 x1, y1, x2, y2（含端点）
+_CYST = (290, 95, 320, 125)
+_BOLUS = LognormalParameters(area=1300.0, mu=2.4, sigma=0.45, arrival_offset_s=4.0, baseline=40.0)
+_INJECTION_OFFSET_S = 5.0
+
+
+def _write_video(path: Path) -> None:
+    rng = np.random.default_rng(3)
+    times = np.arange(_FRAMES) / _FPS
+    intensity = evaluate_lognormal(times, _BOLUS)
+    writer = cv2.VideoWriter(str(path), cv2.VideoWriter.fourcc(*"mp4v"), _FPS, (_WIDTH, _HEIGHT))
+    x1, y1, x2, y2 = _LESION
+    for index in range(_FRAMES):
+        frame = np.zeros((_HEIGHT, _WIDTH, 3), dtype=np.uint8)
+        frame[:, :_SPLIT] = 120
+        frame[:, _SPLIT + 1 :] = 30
+        lesion = np.clip(
+            intensity[index] + rng.normal(0.0, 6.0, (y2 - y1 + 1, x2 - x1 + 1)), 0, 255
+        )
+        frame[y1 : y2 + 1, x1 : x2 + 1] = lesion.astype(np.uint8)[:, :, None]
+        cx1, cy1, cx2, cy2 = _CYST
+        frame[cy1 : cy2 + 1, cx1 : cx2 + 1] = 10
+        cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), thickness=3)
+        writer.write(frame)
+    writer.release()
+
+
+@pytest.fixture(scope="module")
+def config() -> CeusConfig:
+    from conftest import load_ceus_section
+
+    return CeusConfig.model_validate(load_ceus_section())
+
+
+@pytest.fixture(scope="module")
+def prepared(tmp_path_factory: pytest.TempPathFactory, config: CeusConfig) -> SegmentationInput:
+    path = tmp_path_factory.mktemp("ceus") / "synthetic.mp4"
+    _write_video(path)
+    return prepare_segmentation(path, config)
+
+
+def _lesion_masks(prepared: SegmentationInput) -> np.ndarray:
+    """合成「分割结果」：每帧都是病灶方块挪到左半局部坐标后的位置（略往里收 3 像素）。"""
+    view = prepared.contrast.target_view
+    masks = np.zeros((prepared.frame_count, view.height, view.width), dtype=bool)
+    x1, y1, x2, y2 = _LESION
+    masks[:, y1 + 3 : y2 - 2, x1 - _SPLIT + 3 : x2 - _SPLIT - 2] = True
+    return masks
+
+
+def test_prepare_segmentation_finds_contrast_half_and_maps_prompts(
+    prepared: SegmentationInput,
+) -> None:
+    assert prepared.frame_count == _FRAMES
+    assert prepared.frame_size == (_WIDTH, _HEIGHT)
+    assert prepared.contrast.side == "right"
+    target = prepared.contrast.target_view
+    assert (target.x1, target.x2) == (0, prepared.split_x)
+    assert abs(prepared.split_x - _SPLIT) <= 1
+    assert len(prepared.target_frames) == _FRAMES
+    assert all(frame.shape == (target.height, target.width, 3) for frame in prepared.target_frames)
+    assert sorted(prepared.prompt_masks) == [p.frame_index for p in prepared.contrast.prompt_frames]
+    assert 0 in prepared.prompt_masks
+    xs = np.where(prepared.prompt_masks[0])[1]
+    # 描边 3 像素宽、找绿框时再膨胀 2 像素、mp4 压缩会糊开边缘，容差放到 5 像素。
+    assert abs(int(xs.min()) - 50) <= 5
+    assert abs(int(xs.max()) - 169) <= 5
+
+
+def test_quantify_case_end_to_end(prepared: SegmentationInput, config: CeusConfig) -> None:
+    result = quantify_case(prepared, _lesion_masks(prepared), _INJECTION_OFFSET_S, config)
+
+    assert result.trajectory.source_polygons.shape == (_FRAMES, 64, 2)
+    assert result.lesion_fit_ready is True
+    assert list(result.anchors) == list(ANCHOR_SLOTS)
+    for slot in ("D1", "D2", "D3"):
+        assert result.anchors[slot].status == "ok", result.anchors[slot].reason
+    assert result.anchors["F1"].frame_index == 150
+    assert result.anchors["F1"].frame_time_video_s == pytest.approx(15.0, abs=0.06)
+    assert result.anchors["F2"].status == "not_covered"
+    assert result.anchors["F2"].reason is not None
+    assert result.perfusion["roi_perf"] is not None
+    peak_time = _BOLUS.arrival_offset_s + math.exp(_BOLUS.mu - _BOLUS.sigma**2)
+    assert result.perfusion["roi_perf"].ttp_s == pytest.approx(peak_time, abs=0.5)
+    # 画面上没有紫圈。
+    assert result.pan_detected_fraction == 0.0
+    assert result.tic.reference_valid is False
+    # 暗区（囊）被找出来，去暗区曲线比病灶整体亮。
+    peak_row = result.tic.rows[round(peak_time * _FPS)]
+    assert peak_row["dark_area_px"] > 0
+    assert peak_row["roi_perf_raw"] > peak_row["roi_all_raw"]
+
+    restored = unpack_trajectory(pack_trajectory(result.trajectory))
+    assert restored.calibration["contrast_side"] == "right"
+    curves = unpack_tic_curves(pack_tic_curves(result.curves))
+    assert curves.injection_offset_s == pytest.approx(_INJECTION_OFFSET_S)
+
+    summary = quant_summary(result)
+    json.dumps(summary, allow_nan=False)
+    assert summary["anchor_frames"]["F1"]["frame_index"] == 150
+    assert summary["perfusion_metrics"]["roi_perf"] is not None
+
+    markers = tic_plot_markers(result.anchors)
+    assert [marker.label.split()[0] for marker in markers][:3] == ["D1", "D2", "D3"]
+    assert render_tic_png(result.tic, markers=markers).startswith(b"\x89PNG")
+    assert len(render_case_overlay(prepared, result)) > 0
+    images = keyframe_images(prepared, result.anchors)
+    assert "F1" in images and "F2" not in images
+    assert all(image.shape == (_HEIGHT, _WIDTH, 3) for image in images.values())
+
+
+def test_empty_segmentation_fails_case_at_step_4(
+    prepared: SegmentationInput, config: CeusConfig
+) -> None:
+    empty = np.zeros_like(_lesion_masks(prepared))
+    with pytest.raises(NoValidContourError) as raised:
+        quantify_case(prepared, empty, _INJECTION_OFFSET_S, config)
+    assert raised.value.step == 4
+
+
+def test_mask_shape_mismatch_is_rejected(prepared: SegmentationInput, config: CeusConfig) -> None:
+    masks = _lesion_masks(prepared)[:-1]
+    with pytest.raises(ValueError, match="分割掩膜"):
+        quantify_case(prepared, masks, _INJECTION_OFFSET_S, config)